aetherAI—Asia’s leading medical image AI solution provider—has designed an AI-assisted lymph node assessment workflow, validated with clinical experiments, and revealed the improvement of micrometastasis identification sensitivity from 81% to 95%. Based on the results of its clinical experiments, aetherAI GigaPixel algorithm accurately identified metastatic lymph nodes in gastric cancer (AUC, 0.99). A peer-reviewed journal, Nature Communications (Impact IF 2021: 17.69) has published the research. The full article can be accessed online here.
The development of AI applications has been the major driving force for digital transformation of pathology. The extreme resolution of whole slide images, however, posed grave challenges because there is limited memory on the compute accelerator. The typical workaround, the patch-based method, requires detailed annotation by pathologists, which makes the process extremely expensive and slow. aetherAI GigaPixel AI technology was developed to overcome this bottleneck by reducing the memory requirement during neural network computation, thereby allowing AI to be trained using entire undivided whole slide images. This results in dramatically improved throughput of pathology AI development.
“By reducing the need for detailed annotation, aetherAI GigaPixel technology is revolutionizing pathology AI development by training AI models in a more intuitive, economic, and clinically relevant manner,” said Joe Yeh, M.D., aetherAI founder and CEO. “Our research demonstrated the improvement of the pathologists’ review accuracy and decreased their time for review.” The review time per slide was shortened by around 30% on average. “In the near future, it’s expected to see the implementation of such AI-powered workflows in the real world to enhance the diagnosis quality,” Yeh emphasized.
The clinical experiments were conducted by using AI models trained on close to 6000 high-resolution lymph node histopathology images and deployed into daily clinical workflow in Chang Gung Memorial Hospital group. CGMH Group is the largest private hospital group in Taiwan. It has been partnering with aetherAI in creating pathology AI applications to improve diagnostic efficiency since its pathology department underwent digital transformation in 2019.
雲象最新發表「胃癌淋巴結轉移AI偵測」
AI 輔助可縮短病理閱片逾三成時間
雲象科技與林口長庚醫院病理部合作「胃癌淋巴結轉移AI偵測」，並設計 AI 輔助的數位病理工作流程，臨床實驗證明，醫師對胃癌微小轉移病灶的診斷時間縮短了 30%，該演算法辨識率高達 AUC 0.99。這項成果也代表雲象獨創技術「免細節標註全玻片運算」True GigaPixel AI 能力再升級，運用全玻片影像，無切割、免細節標註，直接訓練深度學習模型，突破了深度神經網路硬體加速器上記憶體大小的限制，可加速研發進程。研究成果再度登上國際科學期刊《Nature Communications》，全文可供閱覽。
偕同林口長庚以 50 億畫素影像開發
胃癌手術後，檢查淋巴結是否有受到轉移的癌症侵犯極為重要。長久以來，病理科醫師透過顯微鏡，無其他科技輔助，憑著肉眼與即時記憶，必須在一大片組織裡逐一尋找微小的病灶，做出最終診斷，耗時耗力。為減輕醫師負擔，並驗證 AI 輔助效益，雲象科技與林口長庚醫院病理部陳澤卿主任以及黃士強醫師合作，採用近 6,000 個高達 50 億畫素的淋巴結標註影像，訓練 AI 模型，開發「胃癌淋巴結轉移AI偵測」以及 AI 輔助的數位病理工作流程，協助病理科醫師進行胃癌淋巴結轉移的診斷工作。
以傳統人工作業，病理科醫師對於小於 2mm 的胃癌微小轉移 (micro-metastasis) 病灶的診斷敏感度是 82%，小於 0.2mm 的單獨性腫瘤細胞 (isolated tumor cells) 病灶的診斷敏感度則是 68%。AI 輔助下，醫師對於微小轉移以及單獨性腫瘤細胞的診斷敏感度，則雙雙提升至 96%。同時，對於微小轉移的診斷時間減少 30%，單獨性腫瘤細胞的診斷時間則縮短了 26%，該演算法辨識率高達 AUC 0.99（完美演算法的 AUC 為 1）。
AI 輔助診斷 + AI 數位病理工作流程，驗證輔助效益
林口長庚醫院每年處理病理標本約 12 萬件，達全國之冠，多年來持續推動病理玻片全面數位化，現階段以輔助病理醫師判讀，並提升效率為主要目的，榮獲多項殊榮，包含 2021 年分別獲醫策會頒發「國家醫療品質獎」智慧解決方案組金獎，以及生策會頒發「國家生技醫療品質獎」智慧醫療組銅獎，肯定以人工智慧輔助，有效強化醫院系統性工作流程，達世界級水準，有助提升整體醫療品質。
獨創「免細節標註全玻片運算」再升級
此次發表成功將「免細節標註全玻片運算」True GigaPixel AI 能力再提升，雲象近年致力突破業界一般對影像進行細節標註，並且分割影像區塊的作法；獨創了直接使用未經分割，且未經細節標註的高解析度淋巴結影像，來訓練AI模型。2020 年以來，從使用統一記憶體（unified memory）開始，至「免細節標註全玻片運算」開發，已取得了近 100 倍的加速幅度，使繁重的計算工作流程能更有效率地進行，亦可節省病理科醫師數百小時的標註時間，大幅加速病理AI的開發流程。這些優勢都展現在本項研究得以在一年完成。
雲象科技執行長葉肇元醫師指出，「胃癌淋巴結轉移AI偵測」配合AI輔助數位病理工作流程，以臨床實驗提出了具體效益。從林口長庚醫院備受肯定的「病理全面數位化」實績顯示，導入「數位病理平台」、輔以病理 AI 應用，是邁向數位轉型之路。如今 True GigaPixel AI 可望加速病理 AI 落地，讓率先導入病理 AI 的醫療院所可穩居優勢，有意加入者後來居上。
雲象「免細節標註全玻片運算」True GigaPixel AI，是繼 2021 年與北醫附醫合作，運用在肺癌數位病理影像，今年再接再厲，應用於胃癌的淋巴結轉移診斷，且技術能力再升級，再次獲《Nature Communications》（Impact Factor 2021: 17.69）青睞刊載。透過學術管道檢視研發成果，為受國際肯定、實力驗證的指標之一。自 2019 年雲象發表的國際期刊論文已累計近 20 篇。